kernel - Address excessive stall in pageout during deadlock avoidance
[dragonfly.git] / sys / vm / vm_pageout.c
1 /*
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1994 John S. Dyson
5  * All rights reserved.
6  * Copyright (c) 1994 David Greenman
7  * All rights reserved.
8  *
9  * This code is derived from software contributed to Berkeley by
10  * The Mach Operating System project at Carnegie-Mellon University.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *      This product includes software developed by the University of
23  *      California, Berkeley and its contributors.
24  * 4. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *      from: @(#)vm_pageout.c  7.4 (Berkeley) 5/7/91
41  *
42  *
43  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
44  * All rights reserved.
45  *
46  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
47  *
48  * Permission to use, copy, modify and distribute this software and
49  * its documentation is hereby granted, provided that both the copyright
50  * notice and this permission notice appear in all copies of the
51  * software, derivative works or modified versions, and any portions
52  * thereof, and that both notices appear in supporting documentation.
53  *
54  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57  *
58  * Carnegie Mellon requests users of this software to return to
59  *
60  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
61  *  School of Computer Science
62  *  Carnegie Mellon University
63  *  Pittsburgh PA 15213-3890
64  *
65  * any improvements or extensions that they make and grant Carnegie the
66  * rights to redistribute these changes.
67  *
68  * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $
69  * $DragonFly: src/sys/vm/vm_pageout.c,v 1.36 2008/07/01 02:02:56 dillon Exp $
70  */
71
72 /*
73  *      The proverbial page-out daemon.
74  */
75
76 #include "opt_vm.h"
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
80 #include <sys/proc.h>
81 #include <sys/kthread.h>
82 #include <sys/resourcevar.h>
83 #include <sys/signalvar.h>
84 #include <sys/vnode.h>
85 #include <sys/vmmeter.h>
86 #include <sys/sysctl.h>
87
88 #include <vm/vm.h>
89 #include <vm/vm_param.h>
90 #include <sys/lock.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_pageout.h>
95 #include <vm/vm_pager.h>
96 #include <vm/swap_pager.h>
97 #include <vm/vm_extern.h>
98
99 #include <sys/thread2.h>
100 #include <vm/vm_page2.h>
101
102 /*
103  * System initialization
104  */
105
106 /* the kernel process "vm_pageout"*/
107 static void vm_pageout (void);
108 static int vm_pageout_clean (vm_page_t);
109 static int vm_pageout_scan (int pass);
110 static int vm_pageout_free_page_calc (vm_size_t count);
111 struct thread *pagethread;
112
113 static struct kproc_desc page_kp = {
114         "pagedaemon",
115         vm_pageout,
116         &pagethread
117 };
118 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
119
120 #if !defined(NO_SWAPPING)
121 /* the kernel process "vm_daemon"*/
122 static void vm_daemon (void);
123 static struct   thread *vmthread;
124
125 static struct kproc_desc vm_kp = {
126         "vmdaemon",
127         vm_daemon,
128         &vmthread
129 };
130 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
131 #endif
132
133
134 int vm_pages_needed=0;          /* Event on which pageout daemon sleeps */
135 int vm_pageout_deficit=0;       /* Estimated number of pages deficit */
136 int vm_pageout_pages_needed=0;  /* flag saying that the pageout daemon needs pages */
137
138 #if !defined(NO_SWAPPING)
139 static int vm_pageout_req_swapout;      /* XXX */
140 static int vm_daemon_needed;
141 #endif
142 static int vm_max_launder = 32;
143 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
144 static int vm_pageout_full_stats_interval = 0;
145 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
146 static int defer_swap_pageouts=0;
147 static int disable_swap_pageouts=0;
148
149 #if defined(NO_SWAPPING)
150 static int vm_swap_enabled=0;
151 static int vm_swap_idle_enabled=0;
152 #else
153 static int vm_swap_enabled=1;
154 static int vm_swap_idle_enabled=0;
155 #endif
156
157 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
158         CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
159
160 SYSCTL_INT(_vm, OID_AUTO, max_launder,
161         CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
162
163 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
164         CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
165
166 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
167         CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
168
169 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
170         CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
171
172 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
173         CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
174
175 #if defined(NO_SWAPPING)
176 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
177         CTLFLAG_RD, &vm_swap_enabled, 0, "");
178 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
179         CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
180 #else
181 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
182         CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
183 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
184         CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
185 #endif
186
187 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
188         CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
189
190 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
191         CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
192
193 static int pageout_lock_miss;
194 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
195         CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
196
197 int vm_load;
198 SYSCTL_INT(_vm, OID_AUTO, vm_load,
199         CTLFLAG_RD, &vm_load, 0, "load on the VM system");
200 int vm_load_enable = 1;
201 SYSCTL_INT(_vm, OID_AUTO, vm_load_enable,
202         CTLFLAG_RW, &vm_load_enable, 0, "enable vm_load rate limiting");
203 #ifdef INVARIANTS
204 int vm_load_debug;
205 SYSCTL_INT(_vm, OID_AUTO, vm_load_debug,
206         CTLFLAG_RW, &vm_load_debug, 0, "debug vm_load");
207 #endif
208
209 #define VM_PAGEOUT_PAGE_COUNT 16
210 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
211
212 int vm_page_max_wired;          /* XXX max # of wired pages system-wide */
213
214 #if !defined(NO_SWAPPING)
215 typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int);
216 static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t);
217 static freeer_fcn_t vm_pageout_object_deactivate_pages;
218 static void vm_req_vmdaemon (void);
219 #endif
220 static void vm_pageout_page_stats(void);
221
222 /*
223  * Update vm_load to slow down faulting processes.
224  */
225 void
226 vm_fault_ratecheck(void)
227 {
228         if (vm_pages_needed) {
229                 if (vm_load < 1000)
230                         ++vm_load;
231         } else {
232                 if (vm_load > 0)
233                         --vm_load;
234         }
235 }
236
237 /*
238  * vm_pageout_clean:
239  *
240  * Clean the page and remove it from the laundry.  The page must not be
241  * busy on-call.
242  * 
243  * We set the busy bit to cause potential page faults on this page to
244  * block.  Note the careful timing, however, the busy bit isn't set till
245  * late and we cannot do anything that will mess with the page.
246  */
247
248 static int
249 vm_pageout_clean(vm_page_t m)
250 {
251         vm_object_t object;
252         vm_page_t mc[2*vm_pageout_page_count];
253         int pageout_count;
254         int ib, is, page_base;
255         vm_pindex_t pindex = m->pindex;
256
257         object = m->object;
258
259         /*
260          * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
261          * with the new swapper, but we could have serious problems paging
262          * out other object types if there is insufficient memory.  
263          *
264          * Unfortunately, checking free memory here is far too late, so the
265          * check has been moved up a procedural level.
266          */
267
268         /*
269          * Don't mess with the page if it's busy, held, or special
270          */
271         if ((m->hold_count != 0) ||
272             ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
273                 return 0;
274         }
275
276         mc[vm_pageout_page_count] = m;
277         pageout_count = 1;
278         page_base = vm_pageout_page_count;
279         ib = 1;
280         is = 1;
281
282         /*
283          * Scan object for clusterable pages.
284          *
285          * We can cluster ONLY if: ->> the page is NOT
286          * clean, wired, busy, held, or mapped into a
287          * buffer, and one of the following:
288          * 1) The page is inactive, or a seldom used
289          *    active page.
290          * -or-
291          * 2) we force the issue.
292          *
293          * During heavy mmap/modification loads the pageout
294          * daemon can really fragment the underlying file
295          * due to flushing pages out of order and not trying
296          * align the clusters (which leave sporatic out-of-order
297          * holes).  To solve this problem we do the reverse scan
298          * first and attempt to align our cluster, then do a 
299          * forward scan if room remains.
300          */
301
302 more:
303         while (ib && pageout_count < vm_pageout_page_count) {
304                 vm_page_t p;
305
306                 if (ib > pindex) {
307                         ib = 0;
308                         break;
309                 }
310
311                 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
312                         ib = 0;
313                         break;
314                 }
315                 if (((p->queue - p->pc) == PQ_CACHE) ||
316                     (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
317                         ib = 0;
318                         break;
319                 }
320                 vm_page_test_dirty(p);
321                 if ((p->dirty & p->valid) == 0 ||
322                     p->queue != PQ_INACTIVE ||
323                     p->wire_count != 0 ||       /* may be held by buf cache */
324                     p->hold_count != 0) {       /* may be undergoing I/O */
325                         ib = 0;
326                         break;
327                 }
328                 mc[--page_base] = p;
329                 ++pageout_count;
330                 ++ib;
331                 /*
332                  * alignment boundry, stop here and switch directions.  Do
333                  * not clear ib.
334                  */
335                 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
336                         break;
337         }
338
339         while (pageout_count < vm_pageout_page_count && 
340             pindex + is < object->size) {
341                 vm_page_t p;
342
343                 if ((p = vm_page_lookup(object, pindex + is)) == NULL)
344                         break;
345                 if (((p->queue - p->pc) == PQ_CACHE) ||
346                     (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
347                         break;
348                 }
349                 vm_page_test_dirty(p);
350                 if ((p->dirty & p->valid) == 0 ||
351                     p->queue != PQ_INACTIVE ||
352                     p->wire_count != 0 ||       /* may be held by buf cache */
353                     p->hold_count != 0) {       /* may be undergoing I/O */
354                         break;
355                 }
356                 mc[page_base + pageout_count] = p;
357                 ++pageout_count;
358                 ++is;
359         }
360
361         /*
362          * If we exhausted our forward scan, continue with the reverse scan
363          * when possible, even past a page boundry.  This catches boundry
364          * conditions.
365          */
366         if (ib && pageout_count < vm_pageout_page_count)
367                 goto more;
368
369         /*
370          * we allow reads during pageouts...
371          */
372         return vm_pageout_flush(&mc[page_base], pageout_count, 0);
373 }
374
375 /*
376  * vm_pageout_flush() - launder the given pages
377  *
378  *      The given pages are laundered.  Note that we setup for the start of
379  *      I/O ( i.e. busy the page ), mark it read-only, and bump the object
380  *      reference count all in here rather then in the parent.  If we want
381  *      the parent to do more sophisticated things we may have to change
382  *      the ordering.
383  */
384 int
385 vm_pageout_flush(vm_page_t *mc, int count, int flags)
386 {
387         vm_object_t object;
388         int pageout_status[count];
389         int numpagedout = 0;
390         int i;
391
392         /*
393          * Initiate I/O.  Bump the vm_page_t->busy counter.
394          */
395         for (i = 0; i < count; i++) {
396                 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count));
397                 vm_page_io_start(mc[i]);
398         }
399
400         /*
401          * We must make the pages read-only.  This will also force the
402          * modified bit in the related pmaps to be cleared.  The pager
403          * cannot clear the bit for us since the I/O completion code
404          * typically runs from an interrupt.  The act of making the page
405          * read-only handles the case for us.
406          */
407         for (i = 0; i < count; i++) {
408                 vm_page_protect(mc[i], VM_PROT_READ);
409         }
410
411         object = mc[0]->object;
412         vm_object_pip_add(object, count);
413
414         vm_pager_put_pages(object, mc, count,
415             (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
416             pageout_status);
417
418         for (i = 0; i < count; i++) {
419                 vm_page_t mt = mc[i];
420
421                 switch (pageout_status[i]) {
422                 case VM_PAGER_OK:
423                         numpagedout++;
424                         break;
425                 case VM_PAGER_PEND:
426                         numpagedout++;
427                         break;
428                 case VM_PAGER_BAD:
429                         /*
430                          * Page outside of range of object. Right now we
431                          * essentially lose the changes by pretending it
432                          * worked.
433                          */
434                         pmap_clear_modify(mt);
435                         vm_page_undirty(mt);
436                         break;
437                 case VM_PAGER_ERROR:
438                 case VM_PAGER_FAIL:
439                         /*
440                          * A page typically cannot be paged out when we
441                          * have run out of swap.  We leave the page
442                          * marked inactive and will try to page it out
443                          * again later.
444                          *
445                          * Starvation of the active page list is used to
446                          * determine when the system is massively memory
447                          * starved.
448                          */
449                         break;
450                 case VM_PAGER_AGAIN:
451                         break;
452                 }
453
454                 /*
455                  * If the operation is still going, leave the page busy to
456                  * block all other accesses. Also, leave the paging in
457                  * progress indicator set so that we don't attempt an object
458                  * collapse.
459                  *
460                  * For any pages which have completed synchronously, 
461                  * deactivate the page if we are under a severe deficit.
462                  * Do not try to enter them into the cache, though, they
463                  * might still be read-heavy.
464                  */
465                 if (pageout_status[i] != VM_PAGER_PEND) {
466                         vm_object_pip_wakeup(object);
467                         vm_page_io_finish(mt);
468                         if (vm_page_count_severe())
469                                 vm_page_deactivate(mt);
470 #if 0
471                         if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
472                                 vm_page_protect(mt, VM_PROT_READ);
473 #endif
474                 }
475         }
476         return numpagedout;
477 }
478
479 #if !defined(NO_SWAPPING)
480 /*
481  *      vm_pageout_object_deactivate_pages
482  *
483  *      deactivate enough pages to satisfy the inactive target
484  *      requirements or if vm_page_proc_limit is set, then
485  *      deactivate all of the pages in the object and its
486  *      backing_objects.
487  *
488  *      The object and map must be locked.
489  */
490 static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *);
491
492 static void
493 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object,
494         vm_pindex_t desired, int map_remove_only)
495 {
496         struct rb_vm_page_scan_info info;
497         int remove_mode;
498
499         if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS)
500                 return;
501
502         while (object) {
503                 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
504                         return;
505                 if (object->paging_in_progress)
506                         return;
507
508                 remove_mode = map_remove_only;
509                 if (object->shadow_count > 1)
510                         remove_mode = 1;
511
512                 /*
513                  * scan the objects entire memory queue.  spl protection is
514                  * required to avoid an interrupt unbusy/free race against
515                  * our busy check.
516                  */
517                 crit_enter();
518                 info.limit = remove_mode;
519                 info.map = map;
520                 info.desired = desired;
521                 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
522                                 vm_pageout_object_deactivate_pages_callback,
523                                 &info
524                 );
525                 crit_exit();
526                 object = object->backing_object;
527         }
528 }
529                                         
530 static int
531 vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data)
532 {
533         struct rb_vm_page_scan_info *info = data;
534         int actcount;
535
536         if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) {
537                 return(-1);
538         }
539         mycpu->gd_cnt.v_pdpages++;
540         if (p->wire_count != 0 || p->hold_count != 0 || p->busy != 0 ||
541             (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
542             !pmap_page_exists_quick(vm_map_pmap(info->map), p)) {
543                 return(0);
544         }
545
546         actcount = pmap_ts_referenced(p);
547         if (actcount) {
548                 vm_page_flag_set(p, PG_REFERENCED);
549         } else if (p->flags & PG_REFERENCED) {
550                 actcount = 1;
551         }
552
553         if ((p->queue != PQ_ACTIVE) &&
554                 (p->flags & PG_REFERENCED)) {
555                 vm_page_activate(p);
556                 p->act_count += actcount;
557                 vm_page_flag_clear(p, PG_REFERENCED);
558         } else if (p->queue == PQ_ACTIVE) {
559                 if ((p->flags & PG_REFERENCED) == 0) {
560                         p->act_count -= min(p->act_count, ACT_DECLINE);
561                         if (!info->limit && (vm_pageout_algorithm || (p->act_count == 0))) {
562                                 vm_page_busy(p);
563                                 vm_page_protect(p, VM_PROT_NONE);
564                                 vm_page_wakeup(p);
565                                 vm_page_deactivate(p);
566                         } else {
567                                 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
568                                 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
569                         }
570                 } else {
571                         vm_page_activate(p);
572                         vm_page_flag_clear(p, PG_REFERENCED);
573                         if (p->act_count < (ACT_MAX - ACT_ADVANCE))
574                                 p->act_count += ACT_ADVANCE;
575                         TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
576                         TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
577                 }
578         } else if (p->queue == PQ_INACTIVE) {
579                 vm_page_busy(p);
580                 vm_page_protect(p, VM_PROT_NONE);
581                 vm_page_wakeup(p);
582         }
583         return(0);
584 }
585
586 /*
587  * deactivate some number of pages in a map, try to do it fairly, but
588  * that is really hard to do.
589  */
590 static void
591 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
592 {
593         vm_map_entry_t tmpe;
594         vm_object_t obj, bigobj;
595         int nothingwired;
596
597         if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) {
598                 return;
599         }
600
601         bigobj = NULL;
602         nothingwired = TRUE;
603
604         /*
605          * first, search out the biggest object, and try to free pages from
606          * that.
607          */
608         tmpe = map->header.next;
609         while (tmpe != &map->header) {
610                 switch(tmpe->maptype) {
611                 case VM_MAPTYPE_NORMAL:
612                 case VM_MAPTYPE_VPAGETABLE:
613                         obj = tmpe->object.vm_object;
614                         if ((obj != NULL) && (obj->shadow_count <= 1) &&
615                                 ((bigobj == NULL) ||
616                                  (bigobj->resident_page_count < obj->resident_page_count))) {
617                                 bigobj = obj;
618                         }
619                         break;
620                 default:
621                         break;
622                 }
623                 if (tmpe->wired_count > 0)
624                         nothingwired = FALSE;
625                 tmpe = tmpe->next;
626         }
627
628         if (bigobj)
629                 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
630
631         /*
632          * Next, hunt around for other pages to deactivate.  We actually
633          * do this search sort of wrong -- .text first is not the best idea.
634          */
635         tmpe = map->header.next;
636         while (tmpe != &map->header) {
637                 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
638                         break;
639                 switch(tmpe->maptype) {
640                 case VM_MAPTYPE_NORMAL:
641                 case VM_MAPTYPE_VPAGETABLE:
642                         obj = tmpe->object.vm_object;
643                         if (obj)
644                                 vm_pageout_object_deactivate_pages(map, obj, desired, 0);
645                         break;
646                 default:
647                         break;
648                 }
649                 tmpe = tmpe->next;
650         };
651
652         /*
653          * Remove all mappings if a process is swapped out, this will free page
654          * table pages.
655          */
656         if (desired == 0 && nothingwired)
657                 pmap_remove(vm_map_pmap(map),
658                             VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
659         vm_map_unlock(map);
660 }
661 #endif
662
663 /*
664  * Don't try to be fancy - being fancy can lead to vnode deadlocks.   We
665  * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can
666  * be trivially freed.
667  */
668 void
669 vm_pageout_page_free(vm_page_t m) 
670 {
671         vm_object_t object = m->object;
672         int type = object->type;
673
674         if (type == OBJT_SWAP || type == OBJT_DEFAULT)
675                 vm_object_reference(object);
676         vm_page_busy(m);
677         vm_page_protect(m, VM_PROT_NONE);
678         vm_page_free(m);
679         if (type == OBJT_SWAP || type == OBJT_DEFAULT)
680                 vm_object_deallocate(object);
681 }
682
683 /*
684  * vm_pageout_scan does the dirty work for the pageout daemon.
685  */
686 struct vm_pageout_scan_info {
687         struct proc *bigproc;
688         vm_offset_t bigsize;
689 };
690
691 static int vm_pageout_scan_callback(struct proc *p, void *data);
692
693 static int
694 vm_pageout_scan(int pass)
695 {
696         struct vm_pageout_scan_info info;
697         vm_page_t m, next;
698         struct vm_page marker;
699         struct vnode *vpfailed;         /* warning, allowed to be stale */
700         int maxscan, pcount;
701         int recycle_count;
702         int inactive_shortage, active_shortage;
703         int inactive_original_shortage;
704         vm_object_t object;
705         int actcount;
706         int vnodes_skipped = 0;
707         int maxlaunder;
708
709         /*
710          * Do whatever cleanup that the pmap code can.
711          */
712         pmap_collect();
713
714         /*
715          * Calculate our target for the number of free+cache pages we
716          * want to get to.  This is higher then the number that causes
717          * allocations to stall (severe) in order to provide hysteresis,
718          * and if we don't make it all the way but get to the minimum
719          * we're happy.
720          */
721         inactive_shortage = vm_paging_target() + vm_pageout_deficit;
722         inactive_original_shortage = inactive_shortage;
723         vm_pageout_deficit = 0;
724
725         /*
726          * Initialize our marker
727          */
728         bzero(&marker, sizeof(marker));
729         marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
730         marker.queue = PQ_INACTIVE;
731         marker.wire_count = 1;
732
733         /*
734          * Start scanning the inactive queue for pages we can move to the
735          * cache or free.  The scan will stop when the target is reached or
736          * we have scanned the entire inactive queue.  Note that m->act_count
737          * is not used to form decisions for the inactive queue, only for the
738          * active queue.
739          *
740          * maxlaunder limits the number of dirty pages we flush per scan.
741          * For most systems a smaller value (16 or 32) is more robust under
742          * extreme memory and disk pressure because any unnecessary writes
743          * to disk can result in extreme performance degredation.  However,
744          * systems with excessive dirty pages (especially when MAP_NOSYNC is
745          * used) will die horribly with limited laundering.  If the pageout
746          * daemon cannot clean enough pages in the first pass, we let it go
747          * all out in succeeding passes.
748          */
749         if ((maxlaunder = vm_max_launder) <= 1)
750                 maxlaunder = 1;
751         if (pass)
752                 maxlaunder = 10000;
753
754         /*
755          * We will generally be in a critical section throughout the 
756          * scan, but we can release it temporarily when we are sitting on a
757          * non-busy page without fear.  this is required to prevent an
758          * interrupt from unbusying or freeing a page prior to our busy
759          * check, leaving us on the wrong queue or checking the wrong
760          * page.
761          */
762         crit_enter();
763 rescan0:
764         vpfailed = NULL;
765         maxscan = vmstats.v_inactive_count;
766         for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
767              m != NULL && maxscan-- > 0 && inactive_shortage > 0;
768              m = next
769          ) {
770                 mycpu->gd_cnt.v_pdpages++;
771
772                 /*
773                  * Give interrupts a chance
774                  */
775                 crit_exit();
776                 crit_enter();
777
778                 /*
779                  * It's easier for some of the conditions below to just loop
780                  * and catch queue changes here rather then check everywhere
781                  * else.
782                  */
783                 if (m->queue != PQ_INACTIVE)
784                         goto rescan0;
785                 next = TAILQ_NEXT(m, pageq);
786
787                 /*
788                  * skip marker pages
789                  */
790                 if (m->flags & PG_MARKER)
791                         continue;
792
793                 /*
794                  * A held page may be undergoing I/O, so skip it.
795                  */
796                 if (m->hold_count) {
797                         TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
798                         TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
799                         ++vm_swapcache_inactive_heuristic;
800                         continue;
801                 }
802
803                 /*
804                  * Dont mess with busy pages, keep in the front of the
805                  * queue, most likely are being paged out.
806                  */
807                 if (m->busy || (m->flags & PG_BUSY)) {
808                         continue;
809                 }
810
811                 if (m->object->ref_count == 0) {
812                         /*
813                          * If the object is not being used, we ignore previous 
814                          * references.
815                          */
816                         vm_page_flag_clear(m, PG_REFERENCED);
817                         pmap_clear_reference(m);
818
819                 } else if (((m->flags & PG_REFERENCED) == 0) &&
820                             (actcount = pmap_ts_referenced(m))) {
821                         /*
822                          * Otherwise, if the page has been referenced while 
823                          * in the inactive queue, we bump the "activation
824                          * count" upwards, making it less likely that the
825                          * page will be added back to the inactive queue
826                          * prematurely again.  Here we check the page tables
827                          * (or emulated bits, if any), given the upper level
828                          * VM system not knowing anything about existing 
829                          * references.
830                          */
831                         vm_page_activate(m);
832                         m->act_count += (actcount + ACT_ADVANCE);
833                         continue;
834                 }
835
836                 /*
837                  * If the upper level VM system knows about any page 
838                  * references, we activate the page.  We also set the 
839                  * "activation count" higher than normal so that we will less 
840                  * likely place pages back onto the inactive queue again.
841                  */
842                 if ((m->flags & PG_REFERENCED) != 0) {
843                         vm_page_flag_clear(m, PG_REFERENCED);
844                         actcount = pmap_ts_referenced(m);
845                         vm_page_activate(m);
846                         m->act_count += (actcount + ACT_ADVANCE + 1);
847                         continue;
848                 }
849
850                 /*
851                  * If the upper level VM system doesn't know anything about 
852                  * the page being dirty, we have to check for it again.  As 
853                  * far as the VM code knows, any partially dirty pages are 
854                  * fully dirty.
855                  *
856                  * Pages marked PG_WRITEABLE may be mapped into the user
857                  * address space of a process running on another cpu.  A
858                  * user process (without holding the MP lock) running on
859                  * another cpu may be able to touch the page while we are
860                  * trying to remove it.  vm_page_cache() will handle this
861                  * case for us.
862                  */
863                 if (m->dirty == 0) {
864                         vm_page_test_dirty(m);
865                 } else {
866                         vm_page_dirty(m);
867                 }
868
869                 if (m->valid == 0) {
870                         /*
871                          * Invalid pages can be easily freed
872                          */
873                         vm_pageout_page_free(m);
874                         mycpu->gd_cnt.v_dfree++;
875                         --inactive_shortage;
876                 } else if (m->dirty == 0) {
877                         /*
878                          * Clean pages can be placed onto the cache queue.
879                          * This effectively frees them.
880                          */
881                         vm_page_cache(m);
882                         --inactive_shortage;
883                 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
884                         /*
885                          * Dirty pages need to be paged out, but flushing
886                          * a page is extremely expensive verses freeing
887                          * a clean page.  Rather then artificially limiting
888                          * the number of pages we can flush, we instead give
889                          * dirty pages extra priority on the inactive queue
890                          * by forcing them to be cycled through the queue
891                          * twice before being flushed, after which the 
892                          * (now clean) page will cycle through once more
893                          * before being freed.  This significantly extends
894                          * the thrash point for a heavily loaded machine.
895                          */
896                         vm_page_flag_set(m, PG_WINATCFLS);
897                         TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
898                         TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
899                         ++vm_swapcache_inactive_heuristic;
900                 } else if (maxlaunder > 0) {
901                         /*
902                          * We always want to try to flush some dirty pages if
903                          * we encounter them, to keep the system stable.
904                          * Normally this number is small, but under extreme
905                          * pressure where there are insufficient clean pages
906                          * on the inactive queue, we may have to go all out.
907                          */
908                         int swap_pageouts_ok;
909                         struct vnode *vp = NULL;
910
911                         object = m->object;
912
913                         if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
914                                 swap_pageouts_ok = 1;
915                         } else {
916                                 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
917                                 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
918                                 vm_page_count_min(0));
919                                                                                 
920                         }
921
922                         /*
923                          * We don't bother paging objects that are "dead".  
924                          * Those objects are in a "rundown" state.
925                          */
926                         if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
927                                 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
928                                 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
929                                 ++vm_swapcache_inactive_heuristic;
930                                 continue;
931                         }
932
933                         /*
934                          * The object is already known NOT to be dead.   It
935                          * is possible for the vget() to block the whole
936                          * pageout daemon, but the new low-memory handling
937                          * code should prevent it.
938                          *
939                          * The previous code skipped locked vnodes and, worse,
940                          * reordered pages in the queue.  This results in
941                          * completely non-deterministic operation because,
942                          * quite often, a vm_fault has initiated an I/O and
943                          * is holding a locked vnode at just the point where
944                          * the pageout daemon is woken up.
945                          *
946                          * We can't wait forever for the vnode lock, we might
947                          * deadlock due to a vn_read() getting stuck in
948                          * vm_wait while holding this vnode.  We skip the 
949                          * vnode if we can't get it in a reasonable amount
950                          * of time.
951                          *
952                          * vpfailed is used to (try to) avoid the case where
953                          * a large number of pages are associated with a
954                          * locked vnode, which could cause the pageout daemon
955                          * to stall for an excessive amount of time.
956                          */
957                         if (object->type == OBJT_VNODE) {
958                                 int flags;
959
960                                 vp = object->handle;
961                                 flags = LK_EXCLUSIVE | LK_NOOBJ;
962                                 if (vp == vpfailed)
963                                         flags |= LK_NOWAIT;
964                                 else
965                                         flags |= LK_TIMELOCK;
966                                 if (vget(vp, flags) != 0) {
967                                         vpfailed = vp;
968                                         ++pageout_lock_miss;
969                                         if (object->flags & OBJ_MIGHTBEDIRTY)
970                                                     vnodes_skipped++;
971                                         continue;
972                                 }
973
974                                 /*
975                                  * The page might have been moved to another
976                                  * queue during potential blocking in vget()
977                                  * above.  The page might have been freed and
978                                  * reused for another vnode.  The object might
979                                  * have been reused for another vnode.
980                                  */
981                                 if (m->queue != PQ_INACTIVE ||
982                                     m->object != object ||
983                                     object->handle != vp) {
984                                         if (object->flags & OBJ_MIGHTBEDIRTY)
985                                                 vnodes_skipped++;
986                                         vput(vp);
987                                         continue;
988                                 }
989         
990                                 /*
991                                  * The page may have been busied during the
992                                  * blocking in vput();  We don't move the
993                                  * page back onto the end of the queue so that
994                                  * statistics are more correct if we don't.
995                                  */
996                                 if (m->busy || (m->flags & PG_BUSY)) {
997                                         vput(vp);
998                                         continue;
999                                 }
1000
1001                                 /*
1002                                  * If the page has become held it might
1003                                  * be undergoing I/O, so skip it
1004                                  */
1005                                 if (m->hold_count) {
1006                                         TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1007                                         TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1008                                         ++vm_swapcache_inactive_heuristic;
1009                                         if (object->flags & OBJ_MIGHTBEDIRTY)
1010                                                 vnodes_skipped++;
1011                                         vput(vp);
1012                                         continue;
1013                                 }
1014                         }
1015
1016                         /*
1017                          * If a page is dirty, then it is either being washed
1018                          * (but not yet cleaned) or it is still in the
1019                          * laundry.  If it is still in the laundry, then we
1020                          * start the cleaning operation. 
1021                          *
1022                          * This operation may cluster, invalidating the 'next'
1023                          * pointer.  To prevent an inordinate number of
1024                          * restarts we use our marker to remember our place.
1025                          *
1026                          * decrement inactive_shortage on success to account
1027                          * for the (future) cleaned page.  Otherwise we
1028                          * could wind up laundering or cleaning too many
1029                          * pages.
1030                          */
1031                         TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
1032                         if (vm_pageout_clean(m) != 0) {
1033                                 --inactive_shortage;
1034                                 --maxlaunder;
1035                         }
1036                         next = TAILQ_NEXT(&marker, pageq);
1037                         TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
1038                         if (vp != NULL)
1039                                 vput(vp);
1040                 }
1041         }
1042
1043         /*
1044          * We want to move pages from the active queue to the inactive
1045          * queue to get the inactive queue to the inactive target.  If
1046          * we still have a page shortage from above we try to directly free
1047          * clean pages instead of moving them.
1048          *
1049          * If we do still have a shortage we keep track of the number of
1050          * pages we free or cache (recycle_count) as a measure of thrashing
1051          * between the active and inactive queues.
1052          *
1053          * If we were able to completely satisfy the free+cache targets
1054          * from the inactive pool we limit the number of pages we move
1055          * from the active pool to the inactive pool to 2x the pages we
1056          * had removed from the inactive pool (with a minimum of 1/5 the
1057          * inactive target).  If we were not able to completely satisfy
1058          * the free+cache targets we go for the whole target aggressively.
1059          *
1060          * NOTE: Both variables can end up negative.
1061          * NOTE: We are still in a critical section.
1062          */
1063         active_shortage = vmstats.v_inactive_target - vmstats.v_inactive_count;
1064         if (inactive_original_shortage < vmstats.v_inactive_target / 10)
1065                 inactive_original_shortage = vmstats.v_inactive_target / 10;
1066         if (inactive_shortage <= 0 &&
1067             active_shortage > inactive_original_shortage * 2) {
1068                 active_shortage = inactive_original_shortage * 2;
1069         }
1070
1071         pcount = vmstats.v_active_count;
1072         recycle_count = 0;
1073         m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1074
1075         while ((m != NULL) && (pcount-- > 0) &&
1076                (inactive_shortage > 0 || active_shortage > 0)
1077         ) {
1078                 /*
1079                  * Give interrupts a chance.
1080                  */
1081                 crit_exit();
1082                 crit_enter();
1083
1084                 /*
1085                  * If the page was ripped out from under us, just stop.
1086                  */
1087                 if (m->queue != PQ_ACTIVE)
1088                         break;
1089                 next = TAILQ_NEXT(m, pageq);
1090
1091                 /*
1092                  * Don't deactivate pages that are busy.
1093                  */
1094                 if ((m->busy != 0) ||
1095                     (m->flags & PG_BUSY) ||
1096                     (m->hold_count != 0)) {
1097                         TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1098                         TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1099                         m = next;
1100                         continue;
1101                 }
1102
1103                 /*
1104                  * The count for pagedaemon pages is done after checking the
1105                  * page for eligibility...
1106                  */
1107                 mycpu->gd_cnt.v_pdpages++;
1108
1109                 /*
1110                  * Check to see "how much" the page has been used and clear
1111                  * the tracking access bits.  If the object has no references
1112                  * don't bother paying the expense.
1113                  */
1114                 actcount = 0;
1115                 if (m->object->ref_count != 0) {
1116                         if (m->flags & PG_REFERENCED)
1117                                 ++actcount;
1118                         actcount += pmap_ts_referenced(m);
1119                         if (actcount) {
1120                                 m->act_count += ACT_ADVANCE + actcount;
1121                                 if (m->act_count > ACT_MAX)
1122                                         m->act_count = ACT_MAX;
1123                         }
1124                 }
1125                 vm_page_flag_clear(m, PG_REFERENCED);
1126
1127                 /*
1128                  * actcount is only valid if the object ref_count is non-zero.
1129                  */
1130                 if (actcount && m->object->ref_count != 0) {
1131                         TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1132                         TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1133                 } else {
1134                         m->act_count -= min(m->act_count, ACT_DECLINE);
1135                         if (vm_pageout_algorithm ||
1136                             m->object->ref_count == 0 ||
1137                             m->act_count < pass + 1
1138                         ) {
1139                                 /*
1140                                  * Deactivate the page.  If we had a
1141                                  * shortage from our inactive scan try to
1142                                  * free (cache) the page instead.
1143                                  *
1144                                  * Don't just blindly cache the page if
1145                                  * we do not have a shortage from the
1146                                  * inactive scan, that could lead to
1147                                  * gigabytes being moved.
1148                                  */
1149                                 --active_shortage;
1150                                 if (inactive_shortage > 0 ||
1151                                     m->object->ref_count == 0) {
1152                                         if (inactive_shortage > 0)
1153                                                 ++recycle_count;
1154                                         vm_page_busy(m);
1155                                         vm_page_protect(m, VM_PROT_NONE);
1156                                         vm_page_wakeup(m);
1157                                         if (m->dirty == 0 &&
1158                                             inactive_shortage > 0) {
1159                                                 --inactive_shortage;
1160                                                 vm_page_cache(m);
1161                                         } else {
1162                                                 vm_page_deactivate(m);
1163                                         }
1164                                 } else {
1165                                         vm_page_deactivate(m);
1166                                 }
1167                         } else {
1168                                 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1169                                 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1170                         }
1171                 }
1172                 m = next;
1173         }
1174
1175         /*
1176          * We try to maintain some *really* free pages, this allows interrupt
1177          * code to be guaranteed space.  Since both cache and free queues 
1178          * are considered basically 'free', moving pages from cache to free
1179          * does not effect other calculations.
1180          *
1181          * NOTE: we are still in a critical section.
1182          *
1183          * Pages moved from PQ_CACHE to totally free are not counted in the
1184          * pages_freed counter.
1185          */
1186         while (vmstats.v_free_count < vmstats.v_free_reserved) {
1187                 static int cache_rover = 0;
1188                 m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE);
1189                 if (m == NULL)
1190                         break;
1191                 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || 
1192                     m->busy || 
1193                     m->hold_count || 
1194                     m->wire_count) {
1195 #ifdef INVARIANTS
1196                         kprintf("Warning: busy page %p found in cache\n", m);
1197 #endif
1198                         vm_page_deactivate(m);
1199                         continue;
1200                 }
1201                 KKASSERT((m->flags & PG_MAPPED) == 0);
1202                 KKASSERT(m->dirty == 0);
1203                 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1204                 vm_pageout_page_free(m);
1205                 mycpu->gd_cnt.v_dfree++;
1206         }
1207
1208         crit_exit();
1209
1210 #if !defined(NO_SWAPPING)
1211         /*
1212          * Idle process swapout -- run once per second.
1213          */
1214         if (vm_swap_idle_enabled) {
1215                 static long lsec;
1216                 if (time_second != lsec) {
1217                         vm_pageout_req_swapout |= VM_SWAP_IDLE;
1218                         vm_req_vmdaemon();
1219                         lsec = time_second;
1220                 }
1221         }
1222 #endif
1223                 
1224         /*
1225          * If we didn't get enough free pages, and we have skipped a vnode
1226          * in a writeable object, wakeup the sync daemon.  And kick swapout
1227          * if we did not get enough free pages.
1228          */
1229         if (vm_paging_target() > 0) {
1230                 if (vnodes_skipped && vm_page_count_min(0))
1231                         speedup_syncer();
1232 #if !defined(NO_SWAPPING)
1233                 if (vm_swap_enabled && vm_page_count_target()) {
1234                         vm_req_vmdaemon();
1235                         vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1236                 }
1237 #endif
1238         }
1239
1240         /*
1241          * Handle catastrophic conditions.  Under good conditions we should
1242          * be at the target, well beyond our minimum.  If we could not even
1243          * reach our minimum the system is under heavy stress.
1244          *
1245          * Determine whether we have run out of memory.  This occurs when
1246          * swap_pager_full is TRUE and the only pages left in the page
1247          * queues are dirty.  We will still likely have page shortages.
1248          *
1249          * - swap_pager_full is set if insufficient swap was
1250          *   available to satisfy a requested pageout.
1251          *
1252          * - the inactive queue is bloated (4 x size of active queue),
1253          *   meaning it is unable to get rid of dirty pages and.
1254          *
1255          * - vm_page_count_min() without counting pages recycled from the
1256          *   active queue (recycle_count) means we could not recover
1257          *   enough pages to meet bare minimum needs.  This test only
1258          *   works if the inactive queue is bloated.
1259          *
1260          * - due to a positive inactive_shortage we shifted the remaining
1261          *   dirty pages from the active queue to the inactive queue
1262          *   trying to find clean ones to free.
1263          */
1264         if (swap_pager_full && vm_page_count_min(recycle_count))
1265                 kprintf("Warning: system low on memory+swap!\n");
1266         if (swap_pager_full && vm_page_count_min(recycle_count) &&
1267             vmstats.v_inactive_count > vmstats.v_active_count * 4 &&
1268             inactive_shortage > 0) {
1269                 /*
1270                  * Kill something.
1271                  */
1272                 info.bigproc = NULL;
1273                 info.bigsize = 0;
1274                 allproc_scan(vm_pageout_scan_callback, &info);
1275                 if (info.bigproc != NULL) {
1276                         killproc(info.bigproc, "out of swap space");
1277                         info.bigproc->p_nice = PRIO_MIN;
1278                         info.bigproc->p_usched->resetpriority(
1279                                 FIRST_LWP_IN_PROC(info.bigproc));
1280                         wakeup(&vmstats.v_free_count);
1281                         PRELE(info.bigproc);
1282                 }
1283         }
1284         return(inactive_shortage);
1285 }
1286
1287 static int
1288 vm_pageout_scan_callback(struct proc *p, void *data)
1289 {
1290         struct vm_pageout_scan_info *info = data;
1291         vm_offset_t size;
1292
1293         /*
1294          * Never kill system processes or init.  If we have configured swap
1295          * then try to avoid killing low-numbered pids.
1296          */
1297         if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1298             ((p->p_pid < 48) && (vm_swap_size != 0))) {
1299                 return (0);
1300         }
1301
1302         /*
1303          * if the process is in a non-running type state,
1304          * don't touch it.
1305          */
1306         if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
1307                 return (0);
1308
1309         /*
1310          * Get the approximate process size.  Note that anonymous pages
1311          * with backing swap will be counted twice, but there should not
1312          * be too many such pages due to the stress the VM system is
1313          * under at this point.
1314          */
1315         size = vmspace_anonymous_count(p->p_vmspace) +
1316                 vmspace_swap_count(p->p_vmspace);
1317
1318         /*
1319          * If the this process is bigger than the biggest one
1320          * remember it.
1321          */
1322         if (info->bigsize < size) {
1323                 if (info->bigproc)
1324                         PRELE(info->bigproc);
1325                 PHOLD(p);
1326                 info->bigproc = p;
1327                 info->bigsize = size;
1328         }
1329         return(0);
1330 }
1331
1332 /*
1333  * This routine tries to maintain the pseudo LRU active queue,
1334  * so that during long periods of time where there is no paging,
1335  * that some statistic accumulation still occurs.  This code
1336  * helps the situation where paging just starts to occur.
1337  */
1338 static void
1339 vm_pageout_page_stats(void)
1340 {
1341         vm_page_t m,next;
1342         int pcount,tpcount;             /* Number of pages to check */
1343         static int fullintervalcount = 0;
1344         int page_shortage;
1345
1346         page_shortage = 
1347             (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) -
1348             (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count);
1349
1350         if (page_shortage <= 0)
1351                 return;
1352
1353         crit_enter();
1354
1355         pcount = vmstats.v_active_count;
1356         fullintervalcount += vm_pageout_stats_interval;
1357         if (fullintervalcount < vm_pageout_full_stats_interval) {
1358                 tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count;
1359                 if (pcount > tpcount)
1360                         pcount = tpcount;
1361         } else {
1362                 fullintervalcount = 0;
1363         }
1364
1365         m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1366         while ((m != NULL) && (pcount-- > 0)) {
1367                 int actcount;
1368
1369                 if (m->queue != PQ_ACTIVE) {
1370                         break;
1371                 }
1372
1373                 next = TAILQ_NEXT(m, pageq);
1374                 /*
1375                  * Don't deactivate pages that are busy.
1376                  */
1377                 if ((m->busy != 0) ||
1378                     (m->flags & PG_BUSY) ||
1379                     (m->hold_count != 0)) {
1380                         TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1381                         TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1382                         m = next;
1383                         continue;
1384                 }
1385
1386                 actcount = 0;
1387                 if (m->flags & PG_REFERENCED) {
1388                         vm_page_flag_clear(m, PG_REFERENCED);
1389                         actcount += 1;
1390                 }
1391
1392                 actcount += pmap_ts_referenced(m);
1393                 if (actcount) {
1394                         m->act_count += ACT_ADVANCE + actcount;
1395                         if (m->act_count > ACT_MAX)
1396                                 m->act_count = ACT_MAX;
1397                         TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1398                         TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1399                 } else {
1400                         if (m->act_count == 0) {
1401                                 /*
1402                                  * We turn off page access, so that we have
1403                                  * more accurate RSS stats.  We don't do this
1404                                  * in the normal page deactivation when the
1405                                  * system is loaded VM wise, because the
1406                                  * cost of the large number of page protect
1407                                  * operations would be higher than the value
1408                                  * of doing the operation.
1409                                  */
1410                                 vm_page_busy(m);
1411                                 vm_page_protect(m, VM_PROT_NONE);
1412                                 vm_page_wakeup(m);
1413                                 vm_page_deactivate(m);
1414                         } else {
1415                                 m->act_count -= min(m->act_count, ACT_DECLINE);
1416                                 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1417                                 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1418                         }
1419                 }
1420
1421                 m = next;
1422         }
1423         crit_exit();
1424 }
1425
1426 static int
1427 vm_pageout_free_page_calc(vm_size_t count)
1428 {
1429         if (count < vmstats.v_page_count)
1430                  return 0;
1431         /*
1432          * free_reserved needs to include enough for the largest swap pager
1433          * structures plus enough for any pv_entry structs when paging.
1434          */
1435         if (vmstats.v_page_count > 1024)
1436                 vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200;
1437         else
1438                 vmstats.v_free_min = 4;
1439         vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1440                 vmstats.v_interrupt_free_min;
1441         vmstats.v_free_reserved = vm_pageout_page_count +
1442                 vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1443         vmstats.v_free_severe = vmstats.v_free_min / 2;
1444         vmstats.v_free_min += vmstats.v_free_reserved;
1445         vmstats.v_free_severe += vmstats.v_free_reserved;
1446         return 1;
1447 }
1448
1449
1450 /*
1451  * vm_pageout is the high level pageout daemon.
1452  */
1453 static void
1454 vm_pageout(void)
1455 {
1456         int pass;
1457         int inactive_shortage;
1458
1459         /*
1460          * Initialize some paging parameters.
1461          */
1462         curthread->td_flags |= TDF_SYSTHREAD;
1463
1464         vmstats.v_interrupt_free_min = 2;
1465         if (vmstats.v_page_count < 2000)
1466                 vm_pageout_page_count = 8;
1467
1468         vm_pageout_free_page_calc(vmstats.v_page_count);
1469
1470         /*
1471          * v_free_target and v_cache_min control pageout hysteresis.  Note
1472          * that these are more a measure of the VM cache queue hysteresis
1473          * then the VM free queue.  Specifically, v_free_target is the
1474          * high water mark (free+cache pages).
1475          *
1476          * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1477          * low water mark, while v_free_min is the stop.  v_cache_min must
1478          * be big enough to handle memory needs while the pageout daemon
1479          * is signalled and run to free more pages.
1480          */
1481         if (vmstats.v_free_count > 6144)
1482                 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
1483         else
1484                 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
1485
1486         /*
1487          * NOTE: With the new buffer cache b_act_count we want the default
1488          *       inactive target to be a percentage of available memory.
1489          *
1490          *       The inactive target essentially determines the minimum
1491          *       number of 'temporary' pages capable of caching one-time-use
1492          *       files when the VM system is otherwise full of pages
1493          *       belonging to multi-time-use files or active program data.
1494          *
1495          * NOTE: The inactive target is aggressively persued only if the
1496          *       inactive queue becomes too small.  If the inactive queue
1497          *       is large enough to satisfy page movement to free+cache
1498          *       then it is repopulated more slowly from the active queue.
1499          *       This allows a general inactive_target default to be set.
1500          *
1501          *       There is an issue here for processes which sit mostly idle
1502          *       'overnight', such as sshd, tcsh, and X.  Any movement from
1503          *       the active queue will eventually cause such pages to
1504          *       recycle eventually causing a lot of paging in the morning.
1505          *       To reduce the incidence of this pages cycled out of the
1506          *       buffer cache are moved directly to the inactive queue if
1507          *       they were only used once or twice.
1508          *
1509          *       The vfs.vm_cycle_point sysctl can be used to adjust this.
1510          *       Increasing the value (up to 64) increases the number of
1511          *       buffer recyclements which go directly to the inactive queue.
1512          */
1513         if (vmstats.v_free_count > 2048) {
1514                 vmstats.v_cache_min = vmstats.v_free_target;
1515                 vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1516         } else {
1517                 vmstats.v_cache_min = 0;
1518                 vmstats.v_cache_max = 0;
1519         }
1520         vmstats.v_inactive_target = vmstats.v_free_count / 4;
1521
1522         /* XXX does not really belong here */
1523         if (vm_page_max_wired == 0)
1524                 vm_page_max_wired = vmstats.v_free_count / 3;
1525
1526         if (vm_pageout_stats_max == 0)
1527                 vm_pageout_stats_max = vmstats.v_free_target;
1528
1529         /*
1530          * Set interval in seconds for stats scan.
1531          */
1532         if (vm_pageout_stats_interval == 0)
1533                 vm_pageout_stats_interval = 5;
1534         if (vm_pageout_full_stats_interval == 0)
1535                 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1536         
1537
1538         /*
1539          * Set maximum free per pass
1540          */
1541         if (vm_pageout_stats_free_max == 0)
1542                 vm_pageout_stats_free_max = 5;
1543
1544         swap_pager_swap_init();
1545         pass = 0;
1546
1547         /*
1548          * The pageout daemon is never done, so loop forever.
1549          */
1550         while (TRUE) {
1551                 int error;
1552
1553                 /*
1554                  * Wait for an action request
1555                  */
1556                 crit_enter();
1557                 if (vm_pages_needed == 0) {
1558                         error = tsleep(&vm_pages_needed,
1559                                        0, "psleep",
1560                                        vm_pageout_stats_interval * hz);
1561                         if (error && vm_pages_needed == 0) {
1562                                 vm_pageout_page_stats();
1563                                 continue;
1564                         }
1565                         vm_pages_needed = 1;
1566                 }
1567                 crit_exit();
1568
1569                 /*
1570                  * If we have enough free memory, wakeup waiters.
1571                  * (This is optional here)
1572                  */
1573                 crit_enter();
1574                 if (!vm_page_count_min(0))
1575                         wakeup(&vmstats.v_free_count);
1576                 mycpu->gd_cnt.v_pdwakeups++;
1577                 crit_exit();
1578
1579                 /*
1580                  * Scan for pageout.  Try to avoid thrashing the system
1581                  * with activity.
1582                  */
1583                 inactive_shortage = vm_pageout_scan(pass);
1584                 if (inactive_shortage > 0) {
1585                         ++pass;
1586                         if (swap_pager_full) {
1587                                 /*
1588                                  * Running out of memory, catastrophic back-off
1589                                  * to one-second intervals.
1590                                  */
1591                                 tsleep(&vm_pages_needed, 0, "pdelay", hz);
1592                         } else if (pass < 10 && vm_pages_needed > 1) {
1593                                 /*
1594                                  * Normal operation, additional processes
1595                                  * have already kicked us.  Retry immediately.
1596                                  */
1597                         } else if (pass < 10) {
1598                                 /*
1599                                  * Normal operation, fewer processes.  Delay
1600                                  * a bit but allow wakeups.
1601                                  */
1602                                 vm_pages_needed = 0;
1603                                 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1604                                 vm_pages_needed = 1;
1605                         } else {
1606                                 /*
1607                                  * We've taken too many passes, forced delay.
1608                                  */
1609                                 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1610                         }
1611                 } else {
1612                         /*
1613                          * Interlocked wakeup of waiters (non-optional)
1614                          */
1615                         pass = 0;
1616                         if (vm_pages_needed && !vm_page_count_min(0)) {
1617                                 wakeup(&vmstats.v_free_count);
1618                                 vm_pages_needed = 0;
1619                         }
1620                 }
1621         }
1622 }
1623
1624 /*
1625  * Called after allocating a page out of the cache or free queue
1626  * to possibly wake the pagedaemon up to replentish our supply.
1627  *
1628  * We try to generate some hysteresis by waking the pagedaemon up
1629  * when our free+cache pages go below the severe level.  The pagedaemon
1630  * tries to get the count back up to at least the minimum, and through
1631  * to the target level if possible.
1632  *
1633  * If the pagedaemon is already active bump vm_pages_needed as a hint
1634  * that there are even more requests pending.
1635  */
1636 void
1637 pagedaemon_wakeup(void)
1638 {
1639         if (vm_page_count_severe() && curthread != pagethread) {
1640                 if (vm_pages_needed == 0) {
1641                         vm_pages_needed = 1;
1642                         wakeup(&vm_pages_needed);
1643                 } else if (vm_page_count_min(0)) {
1644                         ++vm_pages_needed;
1645                 }
1646         }
1647 }
1648
1649 #if !defined(NO_SWAPPING)
1650 static void
1651 vm_req_vmdaemon(void)
1652 {
1653         static int lastrun = 0;
1654
1655         if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1656                 wakeup(&vm_daemon_needed);
1657                 lastrun = ticks;
1658         }
1659 }
1660
1661 static int vm_daemon_callback(struct proc *p, void *data __unused);
1662
1663 static void
1664 vm_daemon(void)
1665 {
1666         while (TRUE) {
1667                 tsleep(&vm_daemon_needed, 0, "psleep", 0);
1668                 if (vm_pageout_req_swapout) {
1669                         swapout_procs(vm_pageout_req_swapout);
1670                         vm_pageout_req_swapout = 0;
1671                 }
1672                 /*
1673                  * scan the processes for exceeding their rlimits or if
1674                  * process is swapped out -- deactivate pages
1675                  */
1676                 allproc_scan(vm_daemon_callback, NULL);
1677         }
1678 }
1679
1680 static int
1681 vm_daemon_callback(struct proc *p, void *data __unused)
1682 {
1683         vm_pindex_t limit, size;
1684
1685         /*
1686          * if this is a system process or if we have already
1687          * looked at this process, skip it.
1688          */
1689         if (p->p_flag & (P_SYSTEM | P_WEXIT))
1690                 return (0);
1691
1692         /*
1693          * if the process is in a non-running type state,
1694          * don't touch it.
1695          */
1696         if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
1697                 return (0);
1698
1699         /*
1700          * get a limit
1701          */
1702         limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1703                                 p->p_rlimit[RLIMIT_RSS].rlim_max));
1704
1705         /*
1706          * let processes that are swapped out really be
1707          * swapped out.  Set the limit to nothing to get as
1708          * many pages out to swap as possible.
1709          */
1710         if (p->p_flag & P_SWAPPEDOUT)
1711                 limit = 0;
1712
1713         size = vmspace_resident_count(p->p_vmspace);
1714         if (limit >= 0 && size >= limit) {
1715                 vm_pageout_map_deactivate_pages(
1716                     &p->p_vmspace->vm_map, limit);
1717         }
1718         return (0);
1719 }
1720
1721 #endif